seb314

For running llms _locally_ on Android, there's "pocketpal" (~7tok/s on a pixel 7 pro for some quant of llama 3.2 3B).

(Not sure if it uses ollama though)

The state space of chess is so huge that even a giant training set would be a _very_ sparse sample of the stockfish-computed value function.

So the network still needs to do some impressive generalization in order to „interpolate“ between those samples.

I think so, anyway (didn‘t read the paper but worked on alphazero-like algorithms for a few years)

For comparison, car liability insurance in germany typically has a coverage limit of 50 or 100M € (not sure about potential caps per victim though). There is also a legal minimum of 7.5M.

I believe that for a species, longer generations means slower evolutionary adaptation to environmental changes.

(at least for our bodies, our hardware. Ideas/technology can iterate somewhat independently.) (longer lifespan and short generations would probably have caused resource shortages at various points in the past)

Some providers promise to build new renewable capacity that matches the consumption of their customers (not sure about the timeframe). I'm not aware of any provider that builds the required storage too, though

A v2 feature might be e2e encryption, seems like that should be feasible

Milk with a multi-month shelf live can't be fresh, I believe. The fresh (mildly processed) milk I know turns bad within about a week, the "extended shelf life" stuff that was introduced a couple of years ago (in Germany) lasts for a few weeks. There is also "H Milch" which tastes very differend and can be stored at room temperature for several months.

how would authentication work when you/someone does pick up the phone?

how would solar result in areas beeing uninhabitable for decades (or much longer?) after we stopped using them?

the article doesn't seem to consider cost of hyperparameter optimization prior to the final training...

btw this can be done with a 'proper' domain name too (e.g. at namecheap)

I'd like to prefer DDG...

but recently I noticed what seem to be blind spots on many (maybe less popular?) topics. e.g. I can't find how to setup/configure borgbackup on nixos via ddg for any of ~3-5 query variations I tried. For non-logged-in google with uBlock the relevant documentation is the first hit.

If this concept would work in principle, covid could be reduced to scaling testing capacity. My impression is that estimates about achievable testing capacity don't assume a most-important-short-term-problem-of-mankind priority and resource allocation.

The linked article suggests a novel and much cheaper test, which would be great. But even if that didn't work out, what scale could possibly be feasible with existing tests? Pre-shortage, an RT-PCR seemed to be much cheaper than a missed day of work.